透過高效且可微的陰影估計提升光度立體

Abstract

儘管在各種應用中具有價值，但由於沒有考慮在不同的物體幾何形狀和變化的照明條件下進行精確的陰影估計，傳統的光度立體方法面臨了限制。為了解決這個問題，我們提出了一種基於並行前綴採樣方法的快速而精確的陰影估計演算法，並且使用一個可微分的溫度函數。我們提出的方法可以輕易地用於改進現有的光度立體方法，以更好地估計陰影的結果。 此外，我們進一步通過我們提出的更高階導數損失配置來提高性能。 為了評估我們方法的有效性，我們進行了全面的實驗並將我們的結果與多種無監督和監督方法進行比較。結果表明，我們的方法在平均角誤差（MAE）方面始終優於其他最先進的無監督方法，同時與監督技術保持競爭力。

Although valuable in various applications, traditional photometric stereo approaches have faced limitations due to not considering accurate shadow estimation under different object geometry and varying lighting conditions. We propose a fast and precise shadow estimation algorithm based on a parallel prefix-based sampling method with a differentiable temperature function to address this issue. The proposed method can be easily used to improve existing photometric stereo methods for better estimation of shadow estimation results. In addition, we further improve the performance with our proposed higher-order derivation loss configuration. To assess the effectiveness of our method, we conduct comprehensive experiments and compare our results with diverse unsupervised and supervised approaches. The results demonstrate that our method consistently outperforms other state-of-the-art unsupervised methods regarding mean angular error (MAE) while remaining competitive with supervised techniques.